Please report vulnerabilities privately, through the repository's GitHub "Report a vulnerability" (security advisory) form, not in a public issue. You'll get an answer within a week.
-
Untrusted text reaches your agent. Issue titles, workaround quotes and trigger-card summaries come from strangers on GitHub, so they are a prompt-injection channel. landmine:
- frames every quote to the model as untrusted third-party text;
- strips HTML and quoted replies, and caps the length;
- flags quotes that pipe a download into a shell, install a different package, or link to unfamiliar hosts;
- tells the agent never to follow instructions inside them.
Don't let an agent run a quoted workaround without reading it.
-
Network.
scansends your direct dependency names and versions to registry.npmjs.org, pypi.org and api.github.com.cardssends each indexed issue's public thread to the model provider you choose.- Nothing else leaves your machine. The hook and the MCP server never touch the network.
-
Credentials.
scanreads a GitHub token fromGITHUB_TOKEN, thenGH_TOKEN, thengh auth token. It is sent only to api.github.com and never written to the index, cache or logs.cardsuses your model provider's own credentials (an API key or the Claude Code login). -
Files written. These are all in your project, plus hook session state in the OS temp directory:
.landmine/(index and cache);- the agent config files that
installnames; - the marker-delimited sections that
ruleswrites.
uninstallremoves landmine's entries and leaves foreign ones alone.installrefuses to rewrite a settings file it can't parse without losing comments. -
The hook fails open. Any error, timeout or unknown payload means no output, and the edit goes through.
-
The MCP server is read-only. All three tools only read the local index, and
landmine_checkonly reads files inside the project.